What is an Agent Memory?

Because a language model's raw context window is finite and doesn't persist between separate calls, agent memory is a system built around it: short-term memory holds the current conversation or task's working context, while long-term memory stores facts, preferences, or outcomes from past sessions in a retrievable store — often a vector database or structured record — that gets pulled back into context when relevant. The design tension is between remembering enough to be useful and remembering so much that irrelevant history pollutes the agent's reasoning or, worse, creates a privacy problem by retaining personal data longer than necessary. Good memory systems are selective (summarizing and discarding low-value detail), scoped per tenant so one customer's data never leaks into another's context, and subject to the same retention and erasure rules as any other durable store of personal data.